Write a tool-using agent loop that knows when to stop

Every agent tutorial hides the loop inside somebody else's framework.

BeginnerRead 5 minDoing: 30 to 45 minutes with your API key and an editor open6 steps

▶ Start

What you'll have at the end

A script that runs a bounded loop, calls one real tool, and stops once the model says it's done

Before you start

You need

A script that already sends a plain text question to a live LLM API and gets back a plain text answer, with no tool calling wired in yet.

Not covered

This does not cover what happens when the tool itself throws an error, or capping the loop by dollars spent instead of by a repeat count: those are their own problems.

Leans on

Get the model calling a real tool

Write one real tool and its schema
Send the first message with the tool attached
Run the tool and call the model again

Make the loop stop on its own

Stop the moment there's no tool call left
Cap the repeats so it can't run forever
Run it once and check the whole trace

Where to go next

One bad tool call crashes the whole agent run

for what to do when the real function throws partway through, instead of assuming it always succeeds like this recipe does.

Give a runaway agent a spending cap

for capping the loop by dollars spent rather than by a fixed number of repeats.

Checked 15 Aug 2026

Part of the AI Agents cookbook